lib/Target/NVPTX/NVPTXSubtarget.cpp

   1 //===- NVPTXSubtarget.cpp - NVPTX Subtarget Information -------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements the NVPTX specific subclass of TargetSubtarget.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "NVPTXSubtarget.h"
  15 #include "NVPTXTargetMachine.h"
  16
  17 using namespace llvm;
  18
  19 #define DEBUG_TYPE "nvptx-subtarget"
  20
  21 #define GET_SUBTARGETINFO_ENUM
  22 #define GET_SUBTARGETINFO_TARGET_DESC
  23 #define GET_SUBTARGETINFO_CTOR
  24 #include "NVPTXGenSubtargetInfo.inc"
  25
  26 static cl::opt<bool>
  27     NoF16Math("nvptx-no-f16-math", cl::ZeroOrMore, cl::Hidden,
  28               cl::desc("NVPTX Specific: Disable generation of f16 math ops."),
  29               cl::init(false));
  30
  31 // Pin the vtable to this file.
  32 void NVPTXSubtarget::anchor() {}
  33
  34 NVPTXSubtarget &NVPTXSubtarget::initializeSubtargetDependencies(StringRef CPU,
  35                                                                 StringRef FS) {
  36     // Provide the default CPU if we don't have one.
  37   TargetName = CPU.empty() ? "sm_20" : CPU;
  38
  39   ParseSubtargetFeatures(TargetName, FS);
  40
  41   // Set default to PTX 3.2 (CUDA 5.5)
  42   if (PTXVersion == 0) {
  43     PTXVersion = 32;
  44   }
  45
  46   return *this;
  47 }
  48
  49 NVPTXSubtarget::NVPTXSubtarget(const Triple &TT, const std::string &CPU,
  50                                const std::string &FS,
  51                                const NVPTXTargetMachine &TM)
  52     : NVPTXGenSubtargetInfo(TT, CPU, FS), PTXVersion(0), SmVersion(20), TM(TM),
  53       InstrInfo(), TLInfo(TM, initializeSubtargetDependencies(CPU, FS)),
  54       FrameLowering() {}
  55
  56 bool NVPTXSubtarget::hasImageHandles() const {
  57   // Enable handles for Kepler+, where CUDA supports indirect surfaces and
  58   // textures
  59   if (TM.getDrvInterface() == NVPTX::CUDA)
  60     return (SmVersion >= 30);
  61
  62   // Disabled, otherwise
  63   return false;
  64 }
  65
  66 bool NVPTXSubtarget::allowFP16Math() const {
  67   return hasFP16Math() && NoF16Math == false;
  68 }